DIF Analysis with a Multidimensional Multilevel Mixture Model with Covariates

This study aims to compare observed groups and latent classes, derived from mixture models, in terms of the number of items displaying differential item functioning (DIF) and their corresponding effect size levels. It also seeks to investigate the nature of DIF, utilizing it as a tool for collecting validity evidence in the testing process. In this direction, a dataset comprising 20 items (10 mathematics and 10 science) from the eTIMSS 2019 booklet 1, which included 22 countries, was created. The dataset structure was analyzed and found to exhibit both multidimensional and multilevel characteristics. Results showed that, for both observed group variables, only items with negligible DIF (Level A) were identified, whereas mixture models revealed many items with DIF at effect Levels B and C. Examination of the selected CB2-C2 model provided a more detailed descriptive picture of item-parameter differences across latent classes, particularly with